Bridging the generation gap between text planning and linguistic realization

Abstract

Text planning is traditionally done as a separate process, independent of linguistic realization. However, I show in this paper that it is essential for a text planner to know the linguistic consequences of its actions. The choice of how to realize an element affects bow much information is conveyed (e.g., "Floyd arrived safely in Boston" vs. "Floyd landed safely at Logan") and what other information can be added (e.g., "Peter was deciding for an hour" vs. *"Peter made a decision for an hour"). The choice of realization also affects how the relative salience of the elements being expressed will be perceived (e.g., "The green car is in the garage" vs. "The car in the garage is green"). I have defined an intermediate level of representation for text planning, called the Text Structure. It is an abstract linguistic level that reflects germane linguistic constraints while abstracting away from syntactic detail. This representation allows the text planner to have greater control over the decisions, so that it can take advantage of the expressiveness of language to convey subtleties of meaning. More importantly, the Text Structure allows the generation process overall to be incremental, since it ensures that the text plan being composed will always be expressible in the language.La planification de textes est habituellement rtaliste stpartment, sans tenir compte de la rtalisation linguistique.Cependant, l'auteur dtmontre dans cet article qu'il est essentiel A un planificateur de textes de connaitre les consequences linguistiques de ses actions. Le choix de la mtthode de realisation d'un tltment influe sur la quantite d'informations qui est transmise (par ex.: (( Floyd est arrive sain et sauf A Boston N et (( Floyd a atterri sain et sauf a I'atroportLogan H) et sur quelle autre information peut Ptre ajoutde. Le choix de la mtthode de rtalisation influe tgalement sur la faFon dont les caracttres saillants des tlements qui sortt exprimts seront perps (par ex.: G la voiture verte est dans le garage B et ({ la voiture dans le garage est verte B), L'auteur a dtfini un niveau intermtdiaire de reprtsentation pour la planification de textes qu'il a appelC structure de texte. I1 s'agit d'un niveau linguistique abstrait qui reflete les contraintes linguistiques appropriees tout en s'tloignant du dCtail syntaxique. Cette representation permet au planificateur de textes d'avoir un meilleur contrble des dtcisions, et donc de tirer profit de la force d'expression du langage afin de tenir compte des subtilitts de sens. Plus important encore, la structure de texte permet au processus de generation d'etre incrtmentiel, car elle s'assure que le plan de texte en voie de composition soit toujours exprimable dans le langage.

Refining instructional text generation after evaluation

“…More recently, the limitations of this approach have become apparent [26] and several authors are now investigating how to bridge the gap between the two phases. Methods proposed introduce a phase whose most popular name is sentence planning between the previous two [25,27,35]. In this intermediate phase, 'local' refinement techniques are applied to increase the quality of the generated text.…”

Survey of the State of the Art in Human Language Technology

“…Much fine-scaled text flexibility is thus sacrificed (cf. Meteer, 1991); this also reduces the multilingual effectiveness of such accounts.…”

“…Such data-driven approaches (cf. Hovy, 1988;Kittredge, Korelsky, et al, 1991;Suthers, 1991;Meteer, 1991;McDonald, 1992), allow an improved opportunistic response to the contingencies of particular generation situations. Data-driven critics can be combined with the top-down planning of rhetorical structures in order to improve structures according to aggregation rules (Hovy, 1993) or text heuristics (Scott & de Souza, 1990).…”

Intention, interpretation and the computational structure of language

“…See Reiter and Dale (2000). However, when one explores grammar and logic systematically, in a computational setting, one discovers a plethora of options for content that would identify an object but do not correspond directly to the semantics of any natural utterance (Meteer, 1991;van Deemter, 2002). In general, one cannot transform these descriptions into correct logical forms while working purely at the level of conceptual semantics (Shieber, 1993).…”